A Brief Introduction of Berkeley Deep Drive (BDD)A Brief Introduction of Berkeley Deep Drive (BDD)...

A Brief Introduction of Berkeley Deep Drive (BDD)

Presented by Ching-Yao ChanPATH, UC Berkeley

on behalf of Prof. Trevor Darrell

BDD DirectorEECS, UC Berkeley & PATH Co-Director

• Alpha Go

• Already broadly adopted at many high-tech companies

• A flurry of investments

• A cluster of start-ups

Deep Learning: A Buzzword

• A Type of Machine Learning

• Machine Learning (Artificial Intelligence) systems acquire their knowledge, by extracting patterns from raw data.

• Data representation in deep learning• Mapping representation to output• Deep Learning introduces representation that are

expressed in other simpler representations in multiple layers

• Thus a Multi-Layer Deep Structure.

Deep Learning

Source: http://www.deeplearningbook.org/contents/intro.html

Illustration of a Deep Learning Model

3rd hidden layer: object parts2nd hidden layer: corners and contours1st hidden layer: edgesVisible layer: input pixels

Output: object identity

• Deep Learning dated back to 1940s, known as Cybernetics in 1940s-1960s

• Connectionism in 1980s-1990s

• Current resurgence started in 2006

• In recent years, Deep Learning has advanced significantly due to several contributing factors• Greater computing power (CPU, GPU, Network)• Higher availability an affordability of GPUs• Large collection of data samples available to

train and test algorithms (such as ImageNet, with 14,197,122 images, 21841 synsets indexed)

Deep Learning History

Source: Berkeley Vision and Learning Centerhttp://caffe.berkeleyvision.org/

End-to-End Learning for Vision, Text, Speech

Source: Pieter Abbeel presentation, 04/2016

Source: Pieter Abbeel presentation, 04/2016

Doing Better and Better

Source: Trevor Darrell presentation

• Recent Tesla Incident (May 2016, Florida)

• Supposedly, the Tesla (camera + radar) sensor did not recognize the “side of truck” versus the background sky;

• Can a “deep learning” system recognize an object that is “not the same” as a typical target?

Hypothetically,

Source: Pieter Abbeel presentation, 04/2016

End-to-End Visuomotor

Nvidia demo of Visuomotor Control (April 2016) • End-to-End learning, implementation on Drive-PX2

platform• trained a convolutional neural network (CNN) to map

raw pixels from a single front-facing camera directly to steering commands

• With only human steering inputs as training signals; does not explicitly train the system to detect the outline of roads

• Avoided the needs to recognize human-designated features, such as lane markings, other cars, etc., nor did it include a number of “if-then” rules

• As of 03/2016, 72 hours of training data collected• Use simulation to enhance training• 98% of autonomy (see Nvidia paper) in field testing

State of the Art – Nvidia Example

• Formal and complete safety design verification• Training data• Stability• Credit assignment

• Compliance with functional safety (such as ISO-26262)• safety assurance

• Disruptive proposition of end-to-end solution • Departure from conventional automotive

model

Criticism and Challenges

• Berkeley Vision Learning Center• A consortium that started in 2012• Tremendous advances in computer and deep

learning.• Open-source CAFFE, widely used globally

• Berkeley Deep Drive (BDD) Center• A consortium that started in Spring 2016• Seeking to merge deep learning with

automotive perception and bring computer vision technology to the forefront.

Deep Learning at Berkeley

A Research Alliance

to Investigate State-of-the-Art Technologies

in Computer Vision and Machine Learning

for Automotive Applications

Berkeley Deep Drive

Berkeley Deep Drive

• Current members include:– Audi/VW, Bosch, Ford, Honda, Hyundai, Nvidia,

Panasonic, Qualcomm, Samsung, Toyota– GM, NXP, Sony recently joined– Nexar and Mapillary are contributing partners

• Nexar provides 100,000 hours of driving videos yearly

• Mapillary provides millions of images

• Several more are in agreement reviews

• Sponsors membership• Access to faculty and students• Shared use of research outcome, codes and

data, in repository• Commercial use of BDD software repository,

without further licensing agreement with UC Berkeley

• BDD project and scope of study• Proposals submitted by campus PIs• Sponsors review and rate proposals• Panel consolidates and decides on final

selection of projects to sponsors

Berkeley Deep Drive

• Deep Learning Methodologies• Clockwork FCNs for Fast Video Processing• Cross-modal Transfer Learning• Deep Learning for Tracking• Fast Object Detection and Segmentation• Improving the Scaling of Deep Learning Networks by

Characterizing and Exploiting Soft Convexity• Learning Deep Models Securely on Sensitive Imaging Data with

Cryptographic Gaurantees• Unsupervised Representation Learning for Autonomous Driving

• Deep Learning Implementation• Benchmarks and Leaderboard for Deep Reinforcement Learning• Design Space Exploration for Deep Neural Nets for Advanced

Driver Assistance Systems• FPGA PRET Accelerators of Deep Learning Classifiers for

Autonomous Vehicles• Learning to Drive Under Unstructured Conditions• Low Latency Deep Inference for Self-Driving Vehicles• Secure and Privacy-Preserving Deep Learning

Berkeley Deep Drive Categories of Exemplar Projects, 1 of 2

• Detection and Perception• Pedestrian Models in Urban Environment for Autonomous

Driving and Database of Video Sequences for Model Training, Testing, and Implementation

• Real-Time Perception/Prediction of Traffic Scene with Deep Learning for Autonomous Driving

• Motion Prediction for Urban Autonomous Driving Based on Stochastic Policy Learned via Deep Neural Network

• Inference of Drivers' Intent at Intersections• Outdoor Semantic Scene Segmentation via Multi-modal Sensor

Fusion

• Driver-Vehicle Interaction• Verifiable Control for (Semi)Autonomous Cars that Learn from

Human (Re)Actions• Implicit Communication Through a Car's Motion • Understanding Driver Awareness for Smart Vehicles• Human-Machine Arbitration in Hybrid Driving Systems

Berkeley Deep Drive Categories of Exemplar Projects, 2 of 2

• A progressive emergence

• A worldwide community

• Deep Learning for image and speech recognition widely deployed• Prospects for automated driving, medical

imaging, robots, etc.

• Hardware for embedded applications needed• Nvidia, Qualcomm, etc.

• Probably a way to go for truly intelligent machine

Future of AI/Deep Learning

Questions?Check out BDD website

bdd.Berkeley.edu Thank You!Ching-Yao Chan

California PATH ProgramUniversity of California, Berkeley

cychan@berkeley.edu TEL: 510-665-3621